perf: specialize spawn_fold serving path

[colmugx]

Why

spawn_fold is the dominant tell-oriented actor shape, but it currently delegates to the fully generic command/query loop with Query = Unit and Reply = Unit. Every receipt ask therefore still flows through the generic query callback and tuple-return path.

What changed

• give spawn_fold its own loop and serving path;
• handle Tell(cmd) directly through the command handler;
• handle Ask(()) as a FIFO receipt without invoking a generic query callback or constructing a (State, Reply) pair;
• mark the fold command handler local in the specialized serving loop so MoonBit can inline the function parameter;
• keep the same mailbox, batching, cleanup, cancellation, diagnostics and public API semantics.

Measured result

On the same GitHub runner host (runnervmlun5p), AMD EPYC 9V74, same MoonBit toolchain:

• tell median: 11.914 M/s -> 12.677 M/s (+6.4%)
• handler-cost median: 457.8k/s -> 458.3k/s (flat)
• hotcold-16 P99: 125 us -> 128 us (effectively flat)

All repository tests passed in the experimental benchmark run.

Compiler note

I also tested propagating #locals(on_cmd) one level outward onto fuwaroid_loop_fold. MoonBit rejects that with E4163 (This parameter is expected to be local, but it escapes) because the handler is captured by the handle_cancellation async closure. The locality hint therefore remains on the innermost fuwaroid_serve_fold, where it is valid and benchmarked.